MAST-Pro: Dynamic Mixture-of-Experts for Adaptive Segmentation of Pan-Tumors with Knowledge-Driven Prompts

2025-03-18

AI Summary

Existing tumor segmentation methods struggle to effectively integrate medical prior knowledge, simultaneously capture general anatomical representations and tumor-specific features, and incur prohibitive computational overhead for clinical deployment. To address these challenges, we propose a knowledge-driven, prompt-guided Dynamic Mixture-of-Experts (D-MoE) framework. It leverages anatomical structures and textual descriptions as cross-modal prompts to guide representation learning; introduces a dynamic routing mechanism that adaptively activates expert subnetworks to jointly model universal anatomical priors and tumor heterogeneity; and incorporates parameter-efficient fine-tuning (PEFT) for lightweight adaptation. Evaluated on multi-site tumor datasets, our method achieves a 5.20% average Dice improvement, reduces trainable parameters by 91.04% with no accuracy loss, and significantly lowers deployment cost. This work represents the first integration of knowledge-guided prompting with dynamic MoE architecture, striking a balance between generalizability and clinical practicality.

Technology Category

• Machine Learning: Mixture of Experts (MoE)
• Computer Vision: Multi-modal Vision
• Search and Optimization: Learning to Search

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

Accurate tumor segmentation is crucial for cancer diagnosis and treatment. While foundation models have advanced general-purpose segmentation, existing methods still struggle with: (1) limited incorporation of medical priors, (2) imbalance between generic and tumor-specific features, and (3) high computational costs for clinical adaptation. To address these challenges, we propose MAST-Pro (Mixture-of-experts for Adaptive Segmentation of pan-Tumors with knowledge-driven Prompts), a novel framework that integrates dynamic Mixture-of-Experts (D-MoE) and knowledge-driven prompts for pan-tumor segmentation. Specifically, text and anatomical prompts provide domain-specific priors, guiding tumor representation learning, while D-MoE dynamically selects experts to balance generic and tumor-specific feature learning, improving segmentation accuracy across diverse tumor types. To enhance efficiency, we employ Parameter-Efficient Fine-Tuning (PEFT), optimizing MAST-Pro with significantly reduced computational overhead. Experiments on multi-anatomical tumor datasets demonstrate that MAST-Pro outperforms state-of-the-art approaches, achieving up to a 5.20% improvement in average DSC while reducing trainable parameters by 91.04%, without compromising accuracy.

Problem

• Limited incorporation of medical priors in tumor segmentation.
• Imbalance between generic and tumor-specific features.
• High computational costs for clinical adaptation.

Innovation

• Dynamic Mixture-of-Experts for adaptive segmentation
• Knowledge-driven prompts for tumor-specific learning
• Parameter-Efficient Fine-Tuning reduces computational costs